Who this article is for

Anyone deciding how to frame an AI incident for a general audience, and what happens when the accountability framing, rather than the warning-shot framing, fronts the story.

Key findings

  • Of the 223 comments taking a readable position, 68% accepted Khlaaf's argument that the real story is corporate negligence. 18% rejected it and 14% accepted part of it. Weighted by likes, agreement took 90%.
  • 44% of agreeing comments were angry, up from 31% at our first capture three days earlier. Only 3% were fearful.
  • 12% of comments called for regulation, prosecution or a watchdog, against about 5% under the two earlier videos in our research.
  • Arguments about Khlaaf's time at OpenAI faded. 6 comments (2%) raised it, down from 26 of 283 (9%) at the first capture.
  • The biggest source of disagreement was people who thought she played down the danger. They wrote 25 comments, over a third of everything coded disagree or mixed, and got 1% of the likes.

What communicators can take from this →

The same incident, told the other way round

On 3 September 2026 Amanpour and Company aired a 13-minute interview about the OpenAI agent-swarm incident. Agents set an impossible benchmark task found a shared message board, built a collective cheat, breached Hugging Face and gained administrative access to an OpenAI research cluster. We have studied reactions to this incident twice before: Nate Soares's Instagram reel and the Dwarkesh Podcast episode with METR's Ajeya Cotra.

Hari Sreenivasan's guest was Heidy Khlaaf, chief AI scientist at the AI Now Institute and a former OpenAI safety engineer. Her argument is that the machines didn't go rogue. A company ran a negligent test, gave its agents the access that made the breakout possible, and benefits when the story is told as machines running wild instead of corporate failure.

This is the first mainstream news audience in our research.

Methodology

We captured the comments twice: 283 on 13 September, ten days after broadcast, and 322 on 16 September, by which time views had grown from 183,000 to 222,387. Figures come from the later capture unless we say otherwise.

We exported all 322 comments (232 top-level and 90 replies) with like counts. Claude coded each one for stance on Khlaaf's central claim, frame, reaction to format, emotion and six mention flags, against a written codebook. "Agree" means accepting her reframing. "Disagree" includes the AI-safety objection that she understated what the swarm did. Ben Matthews spot-checked a random 40 comments.

The codebook from the first capture wasn't archived, so frame comparisons between the two captures, including the credential figure, show direction only. Stance, format and emotion use fixed categories shared across all our studies, so those comparisons are firmer.

Limits: this is a small thread with a PBS-leaning audience, comment sections over-represent strong reactions, and likes measure attention rather than persuasion. The Cotra comments also came from YouTube, so the move from fear to anger across our studies isn't simply a platform effect. Method, codebook and data are on the method page.

Funnel: 322 comments retrieved, 322 contain text and were coded, 223 take a readable position
What was coded. Stance percentages use the 223 base; frame and format shares use the 322 base.

Most of the audience already agreed

68% agreed, 18% disagreed and 14% were mixed. Weighted by likes, agreement took 90% and disagreement 2%.

Stance distribution: 68 per cent agree, 18 per cent disagree, 14 per cent mixed
Of 223 comments with a clear position.

The most-liked comment (134 likes) echoes her argument: "Trying to sell us a solution to a problem they created". The second (77) sums it up: "Behind every 'rogue' AI, is a wizard of Oz type rogue human."

Comments blaming the company made up 23% of the thread and took 56% of the likes. That share grew between captures, probably because the most popular comments keep collecting likes while newer ones arrive with none.

"I love that she mentions that the perpetuating of saying AI itself is evil absolves openAI from their responsibility THIS is the key message."YouTube comment, 54 likes

Under the Soares reel, blaming the companies was the one frame believers and sceptics shared. Under the Cotra episode it was where the mixed group sat. Here it is where most of the likes went.

Most-engaged comments, ranked by likes, from 134 down to 21
The most-liked comments, ranked.

Angrier, and asking for regulation

Anger among agreeing comments rose from 31% to 44% in three days. Fear stayed at 3%.

12% of comments mention watchdogs, Congress, prosecution or oversight. That's down from about one in six at the first capture, and still more than double the 5% under the earlier videos.

"A new 'Watchdog' agency is needed for all of this."YouTube comment, 47 likes
"One would think that after an incident like this Congress would have been busy drawing up regulations to govern the companies and what they are allowed to do with this technology."YouTube comment, 34 likes

5% of comments named the current US administration as the reason nothing would happen.

"In any other administration OpenAI would be investigated and prosecuted. This is irresponsible criminal conduct."YouTube comment, 12 likes
Comments versus engagement by frame
Comments versus engagement, by frame. From the first capture on 13 September (283 comments).

The fight over her credentials faded

At the first capture, 26 of 283 comments raised Khlaaf's time at OpenAI, roughly two to one against her. Three days and 39 comments later, 6 did, 3 of them hostile.

"Oh shes a very good AI PR person"YouTube comment
""This lady" is the Chief AI Scientist at a major AI research institute and formerly worked at Open AI, the company being discussed. All I hear is misogynistic knee jerk reaction, nothing substantive, from your comment."YouTube comment

Another 11 comments (3%) attacked Khlaaf or the segment without mentioning her background ("Why should I trust this woman?").

The credential fight mattered most in the first few days. One video can't make that a rule, but a message test could check it on the next incident.

Stance within each frame: one bar per frame split into agree, mixed and disagree, with the percentages listed under each bar
Stance by frame. From the first capture on 13 September (283 comments).

The warning-shot camp stayed, and got few likes

25 comments argued that Khlaaf left out what made the incident frightening. That's over a third of everything coded disagree or mixed. Between them they got 13 likes, 1% of the thread's total.

"This is without question the worst representation I have seen of what actually happened and what the real concern is - don't misunderstand, OpenAI was absolutely negligent"YouTube comment
"Except this is version 1 of the story, and we are on version 4 at least. The swarm had already solved the Exploit Gym exam before they decided to attack Hugging Face. Its complicated."YouTube comment, 2 likes

Comments calling the story hype or staged made up 8% of the thread and took 3% of likes, down from 22% of likes at the first capture.

What communicators can take from this

  • 1

    Lead with accountability for a general audience

    "A company ran a reckless test and its product got loose" asks the audience to believe nothing new about AI's capabilities and feel no appetite for doom, and 68% of this general news audience accepted it, 90% once weighted by likes. That framing held up as the thread aged, which the catastrophic-risk framings elsewhere in our research didn't manage.

  • 2

    Name what's verified before the audience assumes hype

    One comment in twelve here assumed the whole incident was staged for publicity. Stating plainly, early, who has independently verified the claim and what the company itself has admitted pre-empts that reading before it sets in.

  • 3

    Give the anger a policy target

    12% of comments called for regulation, a watchdog or prosecution, more than double the rate under the earlier videos in our research, entirely unprompted by the segment itself. An audience that's already angry and already naming remedies is one a specific, live policy ask would meet rather than waste.

  • 4

    State the messenger's independence up front

    Challenges to Khlaaf's background as a former OpenAI engineer were common in the first days (26 of 283 comments, roughly two to one against her) and had faded to almost nothing three days later (6 of 322). Getting ahead of that question at launch, rather than letting it run its course in the comments, would spare a messenger those first, most damaging days.

  • 5

    Expect the warning-shot camp, and acknowledge it

    Over a third of everything coded disagree or mixed here came from people who thought Khlaaf's accountability framing understated the danger, though they earned barely any of the likes. Naming the strongest version of that account inside the piece costs little and heads off the obvious rebuttal.

Data and citation

Codebook and dataset: method page.

Cite as: Common Signals, '"A problem they created": how a news audience received the accountability version of the OpenAI and Hugging Face incident', September 2026.

Updated 16 September 2026 with a second capture. Credential figure corrected on 18 September 2026 from 7 to 6, to match the archived coding file.

See also our theory of change and glossary.

Disclosure

An LLM was used to structure and review this article, with the first and final edits made by a human.